sun grid engine workshop 2007 - github pagesolesenm.github.io/flex-grid/doc/sge-ws2007-flexlm... ·...

Sun Grid Engine Workshop 2007

Internal/External Resource ManagementInternal/External Resource Management

FlexLM Integration (via load sensors)

Dr. Mark Olesen

Mark.Olesen@emconTechnologies.com

Olesen 2007-09-11 2

SUN GridEngine Workshop 2007FlexLM Integration

Overview

� Company Information

� Background

− Goals

− Terminology

� Standard Resource Management Approaches

− Consumables Only

− Simple Load Sensor

− Combined Load Sensor

� Adjustable Resource Management

− Operation

− Quick Installation

− Remaining Issues

Olesen 2007-09-11 3

Company Information

� OEM emission technology – light and commercial vehicles

• $3 billion business, 19 countries, 7,500 employees

• privately owned – One Equity Partners (JPMorgan Chase & Co)

� Simulation in Augsburg (Europe/Asia Headquarters)

• Acoustics, CFD, FEA

• 60-cpu Linux cluster

• Abaqus, GT-Power, Nastran, OpenFOAM, Star-CD

Olesen 2007-09-11 4

Olesen 2007-09-11 5

Resource Management Goals

� Integrate shared external resource

− eg, software licenses

� Reliable scheduling without collisions

� Give GridEngine maximum control

� Simple installation / maintenance

Olesen 2007-09-11 6

Terminology

� Load Sensors

• qconf -sconf global

− load_sensor pathToLoadsensor

− load_report_time 00:00:60

� Non-Managed Complexes

• qconf -se global

− complex_values NONE

� Managed Complexes

• qconf -se global

− complex_values attr=value[,attr=value,…]

Olesen 2007-09-11 7

Terminology (2)

� Internal Count

• Total number of allocated consumable complexes

• qstat -u \* -xml -r -s prs

<?xml version='1.0'?>

<job_info xmlns:xsd="http://www.w3.org/2001/XMLSchema">

<queue_info>

<job_list state="running">

<JB_job_number>jobNumberjobNumber</JB_job_number>

<JB_owner>userNameuserName</JB_owner>

<state>[prsprs]</state>

<queue_name>queue@host.domainhost.domain</queue_name>

<hard_request name="licenselicense“ …>NN</hard_request>

</job_list>

</queue_info>

</job_info>

Olesen 2007-09-11 8

Standard Resource Management Approaches

� Consumable Resource Management

• Managed consumables

• No load sensor

� Simple Load Sensor Management

• No managed consumables

• Load sensor

� Combined Load Sensor Management

• Managed consumables

• Load sensor

Olesen 2007-09-11 9

Consumable Management

� No load sensor report for this complex

� Managed consumables

• complex_values != NONE

� Resource Availability

= total - internal

� Pros

• Correct internal accounting

• No internal race condition

� Cons

• No external information!

Managed

Internal Count

Availability

Olesen 2007-09-11 10

Simple Load Sensor Management

� Non-managed consumables• complex_values == NONE

� Load sensor report for this complexbeginglobal:resource:Nend

� Resource Availability= reported

� Pros• External information

� Cons• No internal accounting!

Reported

Availability

Olesen 2007-09-11 11

Simple Load Sensor Management (2)

� Possible Race Condition?• “Unfortunately, due to the loadsensor's delay, it can't be 100% excluded that batch jobs

are dispatched and started although the license has been acquired by an interactive job.” †

� Not really race condition … more of a crash!

† http://gridengine.sunsource.net/project/gridengine/howto/resource_management.html

Olesen 2007-09-11 12

Example Crash Condition

� Submit N jobs, with "-l license=4“

• Load sensor: 60 sec

• Scheduler: 5 sec

� t<0 (no licenses)complex_values NONE

(internal_count) NONE

load_values license=0

(available) license=0

� t=0 (licenses become available)complex_values NONE

Olesen 2007-09-11 13

Example Crash Condition (2)

� t=5s (1st interval)

• available license=4

• N jobs start

• N-1 jobs fail

� t=10s (2nd interval)

…complex_values NONE

� t=55s (11th interval)

Olesen 2007-09-11 14

Summary – Simple Load Sensor Management

� Fails immediately

Not useful for any consumable resource!

Olesen 2007-09-11 15

Combined Load Sensor Management

� Load sensor report for this complex

beginglobal:resource:Nend

� Managed consumables

• complex_values != NONE

� Availability

= min( reported, managed – internal count )

Managed

Internal Count

Availability

Reported

Availability

Olesen 2007-09-11 16

Combined Load Sensor Management (2)

� Pros

• Correct internal accounting

• No internal race condition

• Additional external information

� Cons

• None obvious

Olesen 2007-09-11 17

Any Crashes?

� Submit 2 jobs, with "-l license=4“

• Load sensor: 60 sec

� t<0 (no licenses)complex_values license=4

� t=0 (licenses become available)complex_values license=4

Olesen 2007-09-11 18

No Crash!

• available license=4 (but consumable)

• 1 job starts, 1 job waits

� t=10s (2nd interval)complex_values license=4

(internal_count) license=4

� t=60s (report interval)complex_values license=4

Olesen 2007-09-11 19

Really No Crashes?

� t<0 (no licenses)complex_values license=4

� t=0 (some licenses become available)complex_values license=4

Olesen 2007-09-11 20

Fewer Crashes

• available license=2 (but consumable)

• HELP – the 2nd job starts!!

Olesen 2007-09-11 21

Summary – Combined Load Sensor Management

� No crashes if resources are exclusive to GridEngine

• Just use internal resource management instead

Not useful for shared consumable resources!

Olesen 2007-09-11 22

The only solution?

� Let GridEngine do the bookkeeping

• Add some of our own creative bookkeeping

� Don’t use load sensors

• But misuse them if desired

� Reference Implementation ‘qlicserver’

• http://gridengine.sunsource.net/servlets/ProjectDocumentList

• http://gridengine.info/files/qlicserver-1.26.tar.gz

Olesen 2007-09-11 23

The Adjustable Resource Management Approach

� No load sensor report for this complex

� Managed consumables• adjust complex_values on-the-fly

� Resource Availability= managed – internal count

� Pros• More correct accounting

• Fewer race conditions

� Cons• Need some code to adjust values

Managed

Internal Count

Availability

Olesen 2007-09-11 24

� Adjustment of managed complexes

• qconf -mattr exechost complex_values … global

� Classification

eg, FlexLMInformation

Overview – Adjustable Resource Management

Managed

Internal Count

Availability

ClassifyUsage

qstat Info

Cluster External

Cluster Internal

Cluster External

Olesen 2007-09-11 25

Classification – Adjustable Resource Management

� Sample FlexLM output

• lmutil lmstat –a –c $LM_LICENSE_FILE

Users of NASTRAN: (Total of 3 licenses issued; Total of 3 licenses in use)

"NASTRAN" v2006.1130, vendor: MSC

floating license

user1 host1 /dev/tty (v2003.0715) (server/27005 566), start Mon 8/27 18:29

� Extract

• feature user@host nLicenses

� Re-cast FlexLM feature � GridEngine complex

• eg, “GTpowerX” � “gtpower”

Olesen 2007-09-11 26

Classification (2) – Adjustable Resource Management

� What is external usage?

• Any feature/user@host combination not matched inqstat -u \* -xml -r -s prs

� Matching heuristics

• Best match for parallel jobs

− feature/user@host with nLicenses == nSlots

• But also consider “flattened” results

� Managed

= Total – External

Olesen 2007-09-11 27

Adjustment – Adjustable Resource Management

� Query current definitions

• qconf –se exec global

− complex_values … global

� Modify as required

� Implementation in Perl

• Portable. No compilation required

• Regular expressions, threads, etc.

• Quickly modified

• Fast enough

Olesen 2007-09-11 28

Where to bolt this in?

1. Must run periodically• daemon or cron process are okay

2. Failsafe• avoid ‘qconf’ on a stopped cluster

• handle qmaster migration

3. “Startsafe”• Always start with the qmaster

4. Allow some diagnostics (nice to have)• eg, cache results – allows qlic auxiliary program

� Convenient solution• Use a load sensor on the qmaster to call qlicserver

Olesen 2007-09-11 29

Characteristics of qloadsensor

� qmaster host(s) should be admin and exec hosts• eg, add queue with 0 slots

� Load sensor code snippetwhile :

doread input || exit 1

[ "$input" = quit ] && exit 0

echo begin

echo end

# qlicserver runs between load reports

# monitor SGE_CELL/common/act_qmaster to catch migrationif [ “$HOST” = “$SGE_qmaster”]; then

# cache results (for diagnosis)$SGE_site/qlicserver >| $cache_file.TMP 2>> qloadsensor.err[ -s $cache_file.TMP ] && cat $cache_file.TMP >| $cache_file

exit 0

Olesen 2007-09-11 30

Crash Analysis

• Modification interval: 60 sec

� t=0 (licenses become available)complex_values license=4

(external_count) license=0

Olesen 2007-09-11 31

Crash Analysis (2)

� t=0 (some licenses available)complex_values license=2

Olesen 2007-09-11 32

Remaining Race Condition

� t=0 (all licenses available)complex_values license=4

� t=10s – start external job with license=4complex_values license=4

� t=20s – submit job with "-l license=4“

• HELP – the job starts!!

• Check availability in prolog− exit 99 (resubmit)

Olesen 2007-09-11 33

Remaining Race Condition

� t=0 (all licenses available)complex_values license=4

� t=9.5s – start external job with license=1

� t=10s – start job with "-l license=4“

• Prolog check ok

• External job acquired licenses after prolog check

• Race condition

Olesen 2007-09-11 34

Summary – Adjustable Resource Management

� No collisions

� Prolog check avoids many race conditions

� Remaining race condition

• perhaps minor?

Olesen 2007-09-11 35

qlicserver – Quick Install

� Download qlicserver package

• qlicserver –h # -h = help

� Environment settings

• export LM_LICENSE_FILE=…

� Query FlexLM features

• qlicserver -i # -i = info

lookup => { # license features from server

"-GTise" => "gtise", ## reported but unmanaged

"GTpowerX" => "gtpower",

"GTtools" => "gttools", ## new

Olesen 2007-09-11 36

qlicserver – Quick Install (2)

� Edit qlicserver code

• Add ‘lookup’ and other config entries

• No external configuration file – no parse problems

• External ‘limit’ could be useful

� Add new complex definitions

• qlicserver –c # -c = complex

• qconf –mc …

� Initialize managed complexes

• qlicserver -C # -C = complex

Olesen 2007-09-11 37

qlicserver – Quick Install (3)

� Dry-run

• qlicserver -n # ‘-n’ = as per make

� Add into load sensor script

• Normally triggers automatic re-read

• daemon mode at your own risk

Olesen 2007-09-11 38

Problem Diagnosis using qlic

� Check overall usage

• qlic

� Who is using which licenses, where

• qlic –w # -w = who/where

� Query what qhost -F is reporting

• qlic –q # -q = query

� View cache contents

• qlic –d # -d = dump

host = dealog01

age = 0:00:43

feature total limit extern intern wait free

------- ----- ----- ------ ------ ---- ----

abaqus 12 . . 5 . 7

ansys 1 . . * . 1

foam 40 . * . . 40

gtise 4 . 2 * . 2

gtpower 7 4 . . . 4

hexa 2 . . * . 2

med 3 . . * . 3

nastran 4 3 . 3 5 .

proam 4 . . * . 4

prostar 2 . 2 * . .

starcd 28 26 . 16 . 10

starjob 4 . . 1 . 3

starp 24 . . 16 . 8

stars 4 2 . . . 2

tetra 1 . . * . 1

thomat . . . . . .

Olesen 2007-09-11 39

Prolog Safety Net

� Get requested resources• rclist=`qstat -r -j $JOB_ID | sed -ne 's/^.*hard *resource_list: *//p'`

� Check availability• available=`qlicserver –l $rclist,slots=$NSLOTS,JOB_ID=$JOB_ID`

• exitcode=$?

� Exit 99 (re-queue) or other• exit $exitcode

� Caveat

• qstat requires exec host be a submit host!

Olesen 2007-09-11 40

Conclusions

� Adjustable Resource Management

• best method for shared external resources

• avoids most race conditions (with prolog check)

� Remaining race condition

• difficult to avoid

• hopefully infrequent

� qlicserver and qlic implementations

• simple installation, integration and maintenance

• in production use since 2005

Olesen 2007-09-11 41

Possible Future Developments

� Config file for particular options

• eg, limits (in testing)

� Cache qstat queries (in testing)

� Integrate with xml-qstat

• eg, harmonize cache format

� Any new requirements with 6.1 quotas?

� Sharing licenses between clusters

• is a problem

sun grid engine workshop 2007 - github pagesolesenm.github.io/flex-grid/doc/sge-ws2007-flexlm... ·...

Documents

flexlm licensing - nor-par

flexlm end - · pdf fileflexlm end users guide — december...

sun grid engine

ネットワークライセンス管理（flexlm - area...

archived: flexlm end users manual - national instruments

northern arizona wind & sun | off grid & grid-tied solar

flexlm license management - york university...

a beginner's guide to sun grid engine 6 - apache2 ubuntu...

flexlm reference manual · 2008-12-11 · flexlm reference...

sun one grid engine administration and user’s guide ·...

sun grid engine en entornos grid -...

submitting jobs to the sun grid engine

guide d’administration et d’utilisation de sun one grid

flexlm license management -...

1october 9, 2001 sun in scientific & engineering computing...

the rising sun grid

grid computing @ sun€¦ · grid computing @ sun carlo...

ingecon sun® grid support powerfactory dynamic model

pollex v6 installation guide -...

flexlm license management - welcome to freescale - freescale